Real-time Decoding Analyses Of Rat’s Hippocampal And Cortical Ensemble Spikes In Closed-Loop Neural Interfaces

To understand the structure and function of brain,is the most challenging cutting-edge scientific problem in the 21st century.The importance of brain research has been widely accepted all over the world,while huge amount of investment has been attracted to this area Brain-machine interfaces(BMIs)give their users communication and control channels that do not depend on the brain’s normal output channels of peripheral nerves and muscles.The research of transient neural activities has been limited by the limitation of traditional open-loop BMIs,on the contrary,closed-loop BMIs are able to provide causal information.The read out or detection of neural activity events by real-time decoding is an essential requirement for closing the loop.Advances in multielectrode recording devices enable the collection of in vivo ensemble spike activity from neocortical and subcortical circuits,with electrode arrays consisting of hundreds or even thousands of channels.To deal with large quantity of data,scaling and speeding up neural data analysis has become an emerging research topic in neuroscience Many closed-loop experiments require minimal detection latency,typical scenarios include the applications depend on detection of abrupt changes in neuronal activities.To handle the real-time decoding challenge introduced by large scale neural data,A GPU-powered decoding algorithm based on specified two-kernel designs and on-chip memory optimizations is proposed in this thesis.This work significantly speeds up the computation and scales up the data size for real-time processing,in comparison with a CPU implementation,the proposed approach achieves a 20-50 fold speedup in eight tested rat hippocampal,cortical and thalamic ensemble recordings.To realize the online analysis and assessment of hippocampal replay events,based on the proposed parallel algorithm,by accommodating parallel shuffling with less than 15 ms latency for 1000 shuffle samples,the approach in this thesis enables real time assessment of the statistical significance of online-decoded ’memory replay’ candidates during sleepTo speed up the detection of abrupt changes in neuronal ensemble spike activity,based on the Poisson linear dynamical system(PLDS)model,non-Gaussian dynamical noise for modeling a stochastic jump process in the latent state space is introduced in this thesis.To ef-ficiently estimate the state posterior that accommodates non-Gaussian noise and non-Gaussian likelihood,we propose particle filtering and smoothing algorithms for the change-point detec-tion problem,and speed up the computation by using GPU technology.Analysis using both computer simulations and experimental data for acute pain detection shows effectiveness of this approach.On the basis of this change-detection algorithm,the design and implementation of an online BMI system is described in this thesis,which demonstrates the feasibility of the proposed approach.This system manages the tasks such as data acquisition,model training,decoding,real-time display in parallel threads to minimizes the system delay and enables the configuration and management flexibility by custom design.